NVIDIA G80: General Performance Analysis

Great article

 

Thank You

 

You write this as if R5xx doesn't have fully decoupled texturing "threads" - which is rather curious.

Jawed

 

I don't know that you need to go any further than the graph which that comment is analyzing for "what he meant by that".

 

2560x1600 results would be of particular interest to me though it isn't impossible to extrapolate and imagine rough estimates.

 

I miss the fill-rate graphs But I like the "percentage-drop" graphs

It's interesting to see how 4xAA per cycle in G80 really chews through available bandwidth - put another way, all these years of having only 2xAA per cycle were well justified as the bandwidth simply wasn't there.

Jawed

 

Er, no. It was a forward-looking statement: "and thus a competing architecture in the same space". I'm afraid the entire paragraph reads like hyperbole for the sake of it, clearly missing the fact that decoupled texturing is hardly new.

Why would AMD introduce a new GPU without this feature when it's already stamping out two GPU designs with it? Why couch this analysis of a key architectural point of G80 in terms that raise the question? It's just bizarre editorialising, not analysis.

Are there HL2:LC performance figures for AA+AF anywhere in the article? I couldn't find them, I'm a bit confused why HL2:EP1 and HL2:LC results are used in different tests...

Jawed

 

We would have liked to have done 2560x1600 in lieu of 2048x1536, but that's the highest we could go with the available displays.

Have an example of the fillrate graphs you'd like to see?

 

Good job guys. Looking forward to the CUDA article - that's what sets B3D apart from the regular benchmarking crowd. One thing - the introduction mentions the use of synthetic benchmarks yet they don't appear in the article

 

Great stuff from B3d as usual.
Just was curious if you by any chance have compared Vista and XP results particularly in SC: Chaos Theory.
Obviously GF7 has some dodge drivers in Vista, but i found the G80 much better under Vista than in XP so might that make the GTS look a little better if the tests were under Vista?

 

http://www.beyond3d.com/content/reviews/2/16

Though the colours are a bit confusing. The expanding tables are nice there, too

As a side comment, the Rage3D FPS graphs are the best - I'm sure they won't mind if you "steal" the concept

Jawed

 

BTW results were with HDR on, which now I see wasn't on in this performance analysis, my mistake.

 

So a line graph essentially

An example of the graphs you like on Rage3D then?

 

I think you're parsing it differently than I would. That an aside was tossed in mid-stream to remind readers of the features included doesn't necesarily mean that was the action part of the "forward looking" part of the sentence. I took that to be "something similiarly strong", ie. a performance point made in the performance article.

 

Yes, with colours that don't confuse.

http://www.rage3d.com/reviews/video/nvidia8800/index.php?p=9

Without necessarily reducing the count of AA/AF settings being graphed, compared with what B3D is currently doing in the G80 article. Rage3D's clickable settings not only cut the space used on the page which makes understanding easier, but the "animation" is a nice visual shortcut for the trends.

So, similarly animated fillrate graphs for the different settings would be super-groovy

Jawed

 

Completely agree. Love those Rage charts.

 

Oh yeah, the clickable thing. Well I certainly don't mind stealing my own concepts, and it would certainly work well enough here in this particular case, so we'll see.

 

Thinking about it some more, the organisation of colour in the Rage3D FPS graphs is better, too. Colour per GPU on an FPS chart is ok if there's only 2 or 3 GPUs being compared, but I think the Rage3D charts show that using colour for resolution and simply using a consistent ordering for the GPUs being compared, down the Y axis, is what really makes them work so well.

Jawed